Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
IJSRD - International Journal for Scientific Research & Development| Vol. 3, Issue 04, 2015 | ISSN (online): 2321-0613
All rights reserved by www.ijsrd.com 1821
Naïve based Moving Object Detection and Classification in Video using
Background Subtraction
Aseema Mohanty1 Sanjivani Shantaiya
2
1Scholar
2Assistant Professor
1,2DIMAT, CSVTU, Raipur, India
Abstract— The area of real time surveillance is the common
interest of researchers nowadays due to advancement in
technology and thus the amount of work which can be done
in it is still huge. The main objective here was to detect the
moving object in a video and then to classify the
corresponding incoming frames accordingly if it consists of
moving object or not using naïve bayes classifier. This type
of classification helps in frame by frame analysis of the
video and can be used to give better result for classification.
The main focus was on naïve bayes classifier which was
successfully implemented on a video. The performance of
the methodology is found to be 93.64% accurate and has
precision of 93.77% with an error rate of 7.9%.
Key words: Moving object detection, background
subtraction, frame difference, object classification, naïve
bayes classifier, video processing
I. INTRODUCTION
Nowadays moving object detection has become a very
prime area for research due to its use in various computer
vision applications. Beside from the vital benefit of being
able to differentiate video streams into moving and
background content, detecting moving objects provides a
purpose of attention for recognition, classification and
activity scrutiny making these later steps more effective.
Moving object detection is defined as extracting the motion
part from a video stream. It handles separation of moving
objects from the immobile background scene. To recognize
objects of worth in the video sequence and to group together
pixels of these objects is termed as Object Detection. Since
moving objects are normally the principal source of
information, most methods intend on the detection of such
objects. A general tactic for object detection is usage of
information in an individual frame. However, some object
detection methods require usage of the temporal information
computed from a succession of frames to diminish the
number of false detections [1]. The classification of the
pixels in the video stream into either foreground or
background is said as detection of the moving objects in the
video [2].
Moving object detection is the task of identifying
the physical movement of an object in an exceedingly given
region or space. Over previous couple of years, moving
object detection has received abundant of attraction thanks
to its big selection of applications like video surveillance,
human motion analysis, automaton navigation, event
detection, anomaly detection, video conferencing, traffic
analysis and security. Additionally, moving object detection
is extremely important and worthwhile analysis topic in
field of pc vision and video processing since it forms a
crucial step for several complicated processes like video
object classification and video tracking activity.
Subsequently, identification of actual form of moving object
from a given system of video frames becomes relevant.
However, task of detecting actual form of object in motion
becomes tough thanks to numerous challenges like vigorous
scene changes, illumination variations, existence of shadow,
camouflage and bootstrapping drawback [3].
The moving object detection is the elementary step
for most of the applications based on video processing.
Videos are in reality sequel of images, each of which called
is a frame, played in fast enough frequency so that human
eyes can percept the continuousness of its content. It is
evident that all image processing methods can be applied to
separate frames. Moreover, the contents of two successive
frames are generally closely correlated. An image, generally
from a video stream, is distributed into two complementary
groups of pixels. The first group include the pixels which
correspond to foreground objects while the other and
complimentary group include the background pixels. This
result which is the detected object is often showed as a
binary image or as a mask. It is troublesome to mention an
absolute standard regarding what should be recognized as
foreground and what should be distinguished as background
because this description is rather application specific [4].
Background subtraction is a broadly used methodology for
detecting moving objects in videos streams from static
cameras. It is the general way of motion detection. It is a
process that finds the difference of the current image and the
background image to detect the motion region, and it is
commonly efficient to deliver data included object
information. The keynote of this process lies in the
initialization and update of the background image. The
efficiency of both will influence the precision of test results
[5]. It tries to detect moving regions by subtracting the
current image pixel-by-pixel from a reference background
image which is composed by averaging images over time in
an initialization period. The pixels where the variation is
beyond a threshold value are categorized as foreground.
Background subtraction is an extensively used technique for
detecting moving objects in videos from static cameras. The
reasoning behind the method is that of detecting the moving
objects from the variance amid the current frame and a
reference frame, frequently called the “background image”
or “background model”. As a fundamental, the background
image must be a description of the scene without moving
objects and must be kept frequently updated so as to become
accustomed to the unpredictable luminance circumstances
and geometry situations [6].
In visual surveillance system, motion detection is
that the essential and vital step that classifies the moving
foreground object from the background. The segmented
moving foreground object is also humans, vehicles, animals,
flying birds, moving clouds, leaves of a tree, or the other
noise etc. The work of a classification phase within the
visual surveillance system, is to classify the moving
foreground object into predefined categories like single
person, cluster of person, or vehicle, etc. The visual
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1822
surveillance system is generally used for humans and
vehicles. Once the foreground object fit in to the current
category, the latter task like personal identification, object
chase, and activity scrutiny of the detected foreground
object may be done rather more with efficiency and
precisely. There are two main factors which have an effect
on the presentation of object classification. These are image
illustration and classification.
Classification is the task of generalizing familiar
arrangement to relate to new knowledge. it's the matter of
identifying to that of a group of classes (sub-populations) a
new examination goes on the supply of a preparation set of
information contain notes (or instances) whose class
membership is known [7].
The Naive Bayes algorithmic rule may be a
classification algorithm based on the Bayes rule. A naive
Bayes classifier may be a modest probabilistic classifier
grounded on applying Bayes’ theorem along with strong
(naive) independence conventions. It undertakes that the
occurrence (or absence) of a specific characteristic of a class
is distinct to the existence (or absence) of any other
characteristic, specified the class variable. It’s significantly
suited when the dimensionality of the inputs is high.
Parameter estimation for naive Bayes models makes usage
of the tactic of maximum chance. A bonus of the naïve
Bayes classifier is that it solely needs a little quantity of
training information to estimate the parameters essential for
classification [8].
Bayesian classification offers practical learning
algorithms and previous information and observed
information is combined. Bayesian Classification provides a
helpful viewpoint for considerate and calculating several
learning algorithms. It computes explicit probabilities for
hypothesis and it's robust to noise in input data. A naive
Bayes classifier adopts the principle that the presence or
absence of a particular feature is dissimilar to the presence
or absence of the other feature, given the class variable. For
some varieties of probability models, naive Bayes classifiers
is trained very proficiently in a supervised learning situation.
It additionally referred to as idiots’ Bayes, simple Bayes and
independence Bayes [9].
The remaining portion of this paper is systematized
in several sections. Section II describes about the related
works on this topic, Section III explains the proposed
methodology, while the Section 4 consists of the result,
followed by section V on conclusion and the last section VI
is about future scope of the work.
II. RELATED WORKS
Till now a lots of research and work by various researchers
on the moving object detection, background subtraction,
object classification and naïve bayes based had been done.
This is a very enriching and engaging topic and the result
can be implemented in real time system which makes this
area even more attractive for the researchers and with the
advancement in technology lots of works are being done in
it as well as can be introduced in it.
Alper Yilmaz et al. [1] within the survey paper
have presented an in depth survey of object tracking ways
and conjointly provides a brief review of related topics.
They have divided the tracking ways into three classes based
on the utilization of object representations, namely, ways
establishing point correspondence, ways using primitive
geometric models, and ways using contour evolution. They
have enclosed a brief discussion on widespread object
detection ways and have provided elaborate summaries of
object trackers, together with discussion on the object
depictions, parameter estimation schemes used by the
tracking algorithms and the motion models. Abhishek
Kumar Chauhan and Deep Kumar [10] have represented a
video surveillance scenario with real-time moving object
detection and tracking in their work. The detection of
moving object is very important in several tasks, like video
surveillance and moving object pursuit. The planning of a
video surveillance system is focused on automatic
identification of procedures of interest, particularly on
tracking and classification of moving objects. This survey
paper reviews in short analysis works on object detection
and tracking in videos. Mingyang yang [11] in the paper
provides an improved rule based on Background subtraction
and frame differencing algorithm. It uses surendra
background process to acquire background model and
makes usage of frame differencing procedure to update it,
then mines the moving objects from the combination of
images mined from these two strategies. It is presented
through the result that the tactic will satisfy the preciseness
and accuracy within the video sequences while not touching
the need of real-time.
Geetika [12] within the paper has given a brief
discussion of various classification ways that features
decision trees, K-nearest neighbour classifier, naïve Bayes
classifier and neural network. The paper conjointly
discusses some applications of classification model and at
the end the paper is concluded with the brief observations of
these classification models. It shows Fuzzy illustration of
trees add flexibility in information. The nearest-neighbour
methodology suffers severely from what's referred to as the
“curse of dimensionality”. In decision tree little variations
within the training set, the algorithm mat opt to select an
attribute that isn't truly the best one. Naïve Bayes seems to
fare higher than decision Trees because it shows the
importance of all input attributes just in case of heart disease
prediction. Kirubaraj Ragland and P. Tharcis [13] within the
paper conferred different steps involved in tracking objects
in a video sequence, particularly object detection, object
classification and object pursuit. They surveyed the various
strategies out there for detecting, classifying and tracking
objects in a much elaborated manner. The pros and cons of
every of the strategies are mentioned. Object detection
strategies explained are frame differencing, optical flow and
background subtraction. Then, it shows objects could also
be classified based on shape, motion, colour and texture.
Tracking methods involve point based tracking, kernel
based tracking and silhouette based tracking.
Junyan Peng and Patrick P. K. Chan [14] within the
paper proposes the revised Naive Bayes classifier to combat
the focus attack. For every feature within the Naive bayes
classifier, the extra weight based on the amount of ham and
spam containing the feature is added. The weight reduces
the effect of the main focus attack to the features.
Experimental results show that the projected methodology is
stronger underneath the focus attack. The accuracy on the
attacked samples of the projected methodology is beyond
customary Naive Bayes classifier, particularly when the
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1823
degree of attack is massive. Kavitha.A and K.Thulasimani
[15] have worked on object tracking and recognition for real
time videos using naïve Bayes classification. Tracking
methods are categorized on the idea of the object and
motion representations specifically, ways establishing point
correspondence, approaches making usage of primitive
geometric models and procedures using contour evolution.
All these classes need object detection at some point. For
example, the point trackers need detection in each frame,
however geometric region or contours-based trackers
require detection only when the object initial appears within
the scene. Tracking is done based on the detection of the
movable objects, it’s done by Gaussian Mixture Model.
Harris Corner is employed for tracking and the classification
relies on Naive-Bayes Classifier. Naive Bayes provides the
higher classification by its class probability feature.
The motive of this research work is to successfully
implement naïve bayes classifier in video for classifying
each incoming frames into two class one containing the
frames having movable object while the other comprises of
frames without movable object. For this first object is
detected using background subtraction method and then
classification is implemented using naïve bayes classifier.
III. METHODOLOGY
Fig. 1.1: Framework of the Proposed Methodology
The purpose of the proposed methodology is to first detect
moving object from a video and then to classify each
extracted frame into two class from which one comprises of
frames having no moving object while the other comprises
of frames having moving object. The proposed methodology
has two stages namely training and testing as shown in fig.
1.1.
The training process starts with the acquisition of
training video sample and then extracting required frames
from it. In the training phase two types of samples are feed
to the feature extractor each having features very different
from each other so that each belong to different class. The
next step comprises of extracting the features from each of
the extracted frame of the video and then the video sample is
passed to next step where the moving object is detected
from the video using background subtraction method
namely frame difference.
After the moving object is detected that frames
containing only the detected object are made to go through
feature extraction and then all these are features are stored in
training database. Then naïve bayes classifier is trained
according to the inputted features to group the dataset into
the two classes in which one class will have all the features
of the frames with no moving object while the other class
will have all the features of the frames with moving objects.
After this only desirables features are used for classification.
Then in the testing phase the test video is acquired and then
its frames are extracted and then as in training phase the
features of the video are extracted and then the moving
object is detected similarly as in training phase finally the
features are passed into naïve bayes classifier to predict the
result and the output shows the number of frames having
moving object and frames not having moving object in the
train video.
The steps of the proposed methodology comprises as
follows:
1) Video acquisition
2) Pre-processing
3) Moving object detection using background
subtraction
4) Feature extraction from each frame
5) Rank based feature selection
6) Naïve bayes classification
A. Video Acquisition
This step comprises of process of converting the data
produced by a video camera which is obtained in analog
video signal form to digital video. This digital video data
obtained are referred to as video stream or digital video
stream that can be used on computer for further processing
that data.
B. Pre Processing
In this step the acquired data is prepared for processing.
Matlab consists of function VideoReader which is used to
acquire the digital video from the video obtained from a
video camera. Then the read function is used to read each
frame from the video. In the proposed system first 100
frames are extracted from the video and used for analysis
process for the further steps. Then the frames are resized to
fit the system.
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1824
C. Moving Object Detection Using Background
Subtraction
After the first 100 frames are extracted from the acquired
video, moving object detection is performed on it to detect
any type of presence of moving object in that particular
frame. The method used for moving object detection is
frame difference which is a type of background subtraction
method. In this step the efforts are done to extract the
foreground from the background so that to find the moving
object in the video. For this we need a background image
from which the corresponding frames are differentiated so
as to find the objects which are actually moving. The benefit
of using frame difference method is that it is simple and less
complex and gives good result for static background.
In this first the background frame is initialized and
then the subsequent frames are subtracted from it and if the
difference is greater than the defined threshold then it is
taken as foreground thus the moving object else it is taken as
background means the non-moving object. To get better
result after each difference the background is updated with
the current frames which then becomes the background
frame for next frame. Its implementation is as
- Calculate the difference amid pixels of the two frames
and match with a defined threshold value.
|frame Ii – frame Ib|
- If the difference is above threshold value take it as
foreground object otherwise as a background.
If |frame Ii – frame Ib| > T
Then Foreground object
Else
Background object
The result of frame difference can be seen in fig 1.2
in which the first image is the background image, the next is
the current frame and then the next image is the grayscale
image of the current frame and finally in the last image is
the detected moving object.
(a) (b)
(c) (d)
Fig. 1.2: Frame difference method (a ) background frame,
(b) current frame, (c) grayscale frame and (d) Detected
moving object
D. Feature Extraction from Each Frame
In this proposed system two times eight features are
extracted from the same video file frames. The motive here
is to extract sixteen features from the frames so as to classify
them accordingly if they contains moving object or not. For
this two training videos are taken in which one contains
moving objects while another does not contains any moving
objects. If the frames contain no moving object than it
belongs to 0 classes else it belongs to 1 class. In total 16
features are extracted for the same input video. First 8
features are extracted before application of background
subtraction and the rest same 8 features are calculated for
the frames obtained after applying the background
subtraction method. As the features of both data will vary
highly due to difference in image properties for frames
having movable object and the one which does not has it.
Each image has unique properties and the difference
between the background image and current frame is
completely different for frames having moving object and
the frames having non movable object. The features which
are extracted are:
1) Edge Difference
For this feature first the edges are detected in the frames
using canny edge detector. After performing this the
connected components are calculated in the obtained image
and then the difference between the result of background
and current image is taken of connected components. As if
there is no moving object in the frame then the difference
will be very low and if there is moving object present in it
then the difference would be very high.
2) Thinning Difference
This is a morphological operation for skeletonizing the
image so as to detect only the edges in the image. It is
applied to binary image and it produces another binary
image as output. After performing the thinning operation the
connected components are calculated in the obtained image
and then the difference between the result of background
and current image is taken of connected components. If this
value is more than are moving objects in the current frame
else not.
3) Histogram Difference
In this first the histogram is computed of the background
frame and the current frame. If Ij is pixel intensity for jth
frame then for all the pixels in the image first bin is done
means the intensities are divided in small intervals and count
gives the values in each bin. Then the maximum value of
count is found and then the difference between counts of
both image are taken. This difference helps to find the
correlation between the histograms of the frames if this
value is low it means the images are similar thus no movable
object is present else a movable object may be present in the
current frame.
4) Mean Difference
The pixels of the image are calculated. This gives an
average value of pixel intensity. Then the difference of
mean is calculated of current frame and background frame.
If this difference is less it means the images are similar and
it does not contains any moving object while if there is huge
difference it means the images are dissimilar it means there
is a moving object is present in that frame.
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1825
5) Standard Deviation Difference
This feature calculates the standard deviation of all the
pixels present in the image. If the difference is low it means
the images are more or less identical that means the current
frame does not contains a movable object whereas if its
value is high it may contain a movable object.
6) Mean Squared Error (MSE)
It is used to find the quality of an image. It calculates the
cumulative squared error. The difference between MSE
value of background frame image and current frame image
is taken. If the difference value is very high or infinite than
both the images are dissimilar and a movable object is
present in the current frame image and if the difference
value is low then no moving object is present in it.
7) Peak signal to Noise Ratio (PSNR)
It calculates the peak error i.e. the noise between two
images. The difference between PSNR value of background
frame image and current frame image is taken. If the
difference value is very high or infinite than both the images
are similar so no moving object is present in it and if the
difference value is low than a movable object is present in
the current frame image.
8) Structural Similarity (SSIM) Index
It is used to calculate the similarity amongst two images.
Structural info is the knowledge that the pixels have strong
inter-dependencies specifically when they are spatially
adjacent. If it is nearer to 1 then the images are similar and if
it is nearer to 0 then images are not similar.
9) Algorithm for Feature Extraction
Input: video frames
Output: Extracted features
1) Step 1: Background frameB, current frameCi,
i2
2) Step 2: while i <= End Of File
- Bggrayscale(B), Cggrayscale(Ci)
- Detect edge of Bg and Cg using canny detector
- Beconnected components(Bg), Ceconnected
components(Cg)
- EDIFabs ( Be- Ce)
- Convert Bbbinary(Bg), Cbbinary(Cg)
- Thin(Bb), Thin(Cb)
- Btconnected components(Thin(Bg)),
Ctconnected components(Thin(Cg))
- TDIF abs ( Bt- Ct)
- Bh Imhist(Bg), Ch Imhist(Cg)
- HDIFAbs (max (Bh (count)) - max (Ch (count)))
- MDIFAbs (mean (Bg) - mean (Cg))
- STDDIFAbs (Standard_deviation (Bg)-
Standard_deviation (Cg))
- MSDIFMSE (Bg, Cg)
- PSDIFPSNR (Bg, Cg)
- SDIFSSIM (Bg, Cg)
- BCi
- Ii + 1
3) Step 3: Save all the features in the training database
Fig 1.3 (a) and (b) shows the values for extracted
features for two train class video samples for 10 frames.
(a)
(b)
Fig. 1.3: (a) Extracted features for video having no moving object and (b) Extracted features for video having moving object
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1826
E. Rank Based Feature Selection
After the features are extracted, the system is trained and
tested for various input combinations of the extracted
features and the features which gives the best result are
selected finally. An object can contain numerous features
but it is required to choose the most desirable features from
them so as to differentiate well during classification. The
motive of this step is only to get the most optimal result for
classification. For this different combinations of features
which are extracted are tested to get results and then the
features which gives best result are selected while the ones
which fluctuates the result are eliminated. In this ranking of
extracted features are done means features are ranked
according to their effect on classification. The feature giving
best result is ranked first while the one not giving good
result are ranked lower. The purpose of this step is to
discard irrelevant parameters to give the best result.
After the extraction of features for both training
video set difference of the value is calculated as shown in
fig 1.4. From the difference calculated for the extracted
features of both input values it can be seen that some of the
values are zeros. This features thus will have no influence
on deciding the outcome but can fluctuate the result thus
only 12 features from the set of 16 features are selected and
used for object classification using naïve bayes. The features
which are not selected are PSNR and Histogram difference.
Fig. 1.4: Difference of extracted features for both class input videos
F. Naïve Bayes Classification:
The naïve bayes classifier comprises of two stages namely
training and testing. In training the features are feed to the
classifier and the classes to which they belong are
aforementioned. In the testing stage the testing data is
provided to the classifier and it gives the class of the
features given as input. The better the features and their
value the better outcome can be observed. Using the training
features it estimates the probability distribution of that
feature belongs to a particular class. For testing it calculates
the class having largest probability distribution for a feature.
During test data it calculates the probability distribution of
the feature belonging to which class and henceforth gives
the class of the given input data. It assumes class conditional
independence while finding outcome. In this work in total
16 features are extracted from an object out of which only
12 were feasible and 4 were not feasible. Thus the classifier
was trained on its basis and classification was done in two
classes one containing frames with movable object and other
containing frames with non-movable object. The class
having frames with moving object are given outcome 1
while the one not having movable object are given outcome
0. Naïve Bayes classifier predicts X belongs to Class Ci iff
(
) (
) For 1<= j <= m, j <> I (1)
Where (
)
(
)
(2)
- Where (
): Probability of given X
- : Prior probability of hypothesis
- (
): Probability of X given
- : Prior probability of training data X.
With many attributes, the probability of whole
belonging to which class is calculated as:
(
) ∏ (
)
(3)
Which means (
) (
) (
) (
) (4)
IV. RESULTS AND DISCUSSIONS
Fig. 1.5: Comparison of edge difference of frames with
moving object and nonmoving object before background
subtraction
12 features are used for classification and the analysis for
different features for the different class videos are shown in
below graphs. The 12 features for the videos before
application of background subtraction and after application
of background subtraction are studied from figure 1.5 to
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1827
figure 1.16. For the proposed work lots of data have been
used some are globally available as specified in [16, 17, 18,
19] while some were recorded by us.
From figure 1.5 and 1.6 it can be concluded that
edge difference among the frames not having moving object
have less edge difference from the background frame as they
have almost similar features while the frames having
moving object differ largely from the background frame in
terms of edge difference.
Fig. 1.6: Comparison of edge difference of frames with
moving object and nonmoving object after background
subtraction
From figure 1.7 and 1.8 it can be concluded that
thinning difference among the frames not having moving
object have less thinning difference from the background
frame as they have almost similar features while the frames
having moving object differ largely from the background
frame in terms of thinning difference.
Fig. 1.7: Comparison of thinning difference of frames with
moving object and nonmoving object before background
subtraction
Fig. 1.8: Comparison of thinning difference of frames with
moving object and nonmoving object after background
subtraction
From figure 1.9 and 1.10 it can be concluded that MSE
among the frames not having moving object and background
frame is less as they have almost similar features while the
frames having moving object differ largely from the
background frame in terms of MSE.
Fig. 1.9: Comparison of MSE of frames with moving object
and nonmoving object before background subtraction
Fig. 1.10: Comparison of MSE of frames with moving
object and nonmoving object after background subtraction
From figure 1.11 and 1.12 it can be concluded that
SSIM index among the frames not having moving object
and background frame is nearly equal to 1 as they have
almost similar features while the frames having moving
object thus differ largely from the background frame in
terms of SSIM index and is not nearly equal to 1.
Fig. 1.11: Comparison of SSIM index of frames with
moving object and nonmoving object before background
subtraction
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1828
Fig. 1.12: Comparison of SSIM index of frames with
moving object and nonmoving object after background
subtraction
From figure 1.13 and 1.14 it can be concluded that
standard deviation difference among the frames not having
moving object have less standard deviation difference from
the background frame as they have almost similar features
while the frames having moving object differ largely from
the background frame in terms of standard deviation
difference.
Fig. 1.13: Comparison of standard deviation difference of
frames with moving object and nonmoving object before
background subtraction
Fig. 1.14: Comparison of standard deviation difference of
frames with moving object and nonmoving object after
background subtraction
From figure 1.15 and 1.16 it can be concluded that
mean difference among the frames not having moving
object have less mean difference from the background frame
as they have almost similar features while the frames having
moving object differ largely from the background frame in
terms of mean difference.
Fig. 1.15: Comparison of mean difference of frames with
moving object and nonmoving object before background
subtraction
Fig. 1.16: Comparison of mean difference of frames with
moving object and nonmoving object after background
subtraction
After the analysis the proposed methodology is
evaluated as shown in table 1.1.
Videos Precisio
n %
Accurac
y %
Sensitivit
y %
Specificit
y %
Erro
r
rate
%
Video I 100 95.95 95.95 Indef. 4.04
Video
II 100 97.97 97.97 Indef. 2.02
Video
III 100 100 100 Indef. 0
Video
IV 87.5 89.89 100 65.51 10.10
Video
V 91.35 92.92 100 72 7.07
Video
VI 80 89.89 88.88 100 10.10
Video
VII 97.59 88.89 90 77.77 11.11
Averag
e 93.77 93.64 96.11 78.82 7.9
Table 1.1: Performance metrics of proposed methodology
The total run time of each sample video on which
operation is performed is 20 Sec. The total frames which
were extracted are 100. Finally the total run time of the
proposed methodology starting from loading of video to
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1829
classification is on average 15 – 20 minutes. The run time
depends on the quality of the video. If the resolution and
frame per second are less it takes less time but with the
increase in this parameter the running time of the whole
proposed work also increases. The performance of proposed
methodology on the basis of above metrics is given as
93.64% accurate and has precision of 93.77%.
V. CONCLUSION
This research work is based on moving object detection
using background subtraction and then classification of
frame on the basis of them having moving object or not by
using naïve bayes classifier. Firstly for moving object
detection it was observed that frame difference gave an
optimal result in comparison of other background
subtraction techniques as well as it not complex and is also
gives the result very fast. The only drawback in it is it needs
a static background and less illumination and scene changes.
Then the main focus was on naïve bayes classifier which
was successfully implemented on a video. The motive was
to classify the incoming frames from the video into two
class from which one consists of frames with moving
objects while the other contains the frames with no moving
object. The advantages provided by naïve bayes classifier
was the reason it was implemented in the work.
Even though it is a bit time taking the performance
of the methodology is given as 93.64% accurate and has
precision of 93.77% with an error rate of 7.9%.
This work can be applied in many real time
application as now a days video surveillance is a very
common area of interest for researchers and also there is
requirement of surveillance system in today’s era. The
various application included are security surveillance, traffic
monitoring, lane supervision, tracking, video processing,
medical imaging etc.
VI. FUTURE SCOPE
This work can be extended or modified to produce better
results or to perform some other functions. Firstly here
frame difference method is used for moving object detection
which can be replaced by other background subtraction
techniques like Mixture of Gaussian or median filter and the
comparative performance analysis can be done to choose the
better one. Methods like optical flow, segmentation can also
be implemented for moving object detection. Secondly for
classification some other unique properties of an image can
be taken like brightness, entropy, auto correlation,
similarity, pixel quality etc. Also the feature extraction
before application of background subtraction is slow thus
some mechanism could be used to faster the step to get
faster result. It can be used to give better result by removing
noise from the acquired data and also implementing shadow
removal can give much nicer outcome. It can be further
analyzed for multi class classifier and also if number of
moving objects on each frame containing movable objects
can be counted it will give even more benefit.
REFERENCES
[1] Alper Yilmaz, Omar Javed and Mubarak Shah, “Object
Tracking: A Survey” ACM Computing Surveys, Vol.
38, No. 4, Article 13, December 2006.
[2] A.Ramya and Dr. P.Raviraj, “A Survey and
Comparative Analysis of Moving Object Detection and
Tracking”, International Journal of Engineering
Research & Technology, Vol. 2 Issue 10, October –
2013.
[3] Jaya S. Kulchandani and Kruti J. Dangarwala, “Moving
Object Detection: Review of Recent Research Trends”,
Proceedings of IEEE International Conference on
Pervasive Computing (ICPC), -1-4799-6272-
3/15/$31.00(c) 2015.
[4] R. T. Collins et al, “A system for video surveillance and
monitoring: VSAM final report”, Technical report
CMU-RI-TR-00-12, Robotics Institute, Carnegie
Mellon University, May 2000.
[5] Rupali S.Rakibe and Bharati D.Patil, “Background
Subtraction Algorithm Based Human Motion
Detection”, International Journal of Scientific and
Research Publications, Volume 3, Issue 5, May 2013.
[6] M. Piccardi, “Background subtraction techniques: a
review”, in Proceedings IEEE International Conference
Systems, Man, Cybernetics, 2004, pp. 3099–3104.
[7] M.Soundarya, R.Balakrishnan, “Survey on
Classification Techniques in Data mining”,
International Journal of Advanced Research in
Computer and Communication Engineering Vol. 3,
Issue 7, July 2014.
[8] Neha Mehra and Surendra Gupta, “Survey on
Multiclass Classification Methods”, International
Journal of Computer Science and Information
Technologies, Vol. 4 (4), 2013, 572 – 576.
[9] S.Archana and Dr. K.Elangovan, “Survey of
Classification Techniques in Data Mining”,
International Journal of Computer Science and Mobile
Applications, Vol.2 Issue. 2, February- 2014, pg. 65-71.
[10] Abhishek Kumar Chauhan and Prashant Krishan,
“Moving Object Tracking using Gaussian Mixture
Model and Optical Flow”, International Journal of
Advanced Research in Computer Science and Software
Engineering, Volume 3, Issue 4, April 2013.
[11] Mingyang Yang, “Moving Objects Detection Algorithm
in Video Sequence”, 978-1-4799-3903-9/14/$31.00
©2014 IEEE.
[12] Geetika, “A Survey of Classification Methods and its
Applications”, International Journal of Computer
Applications (0975 – 8887) Volume 53– No.17,
September 2012.
[13] Kirubaraj Ragland and P. Tharcis, “A Survey on Object
Detection, Classification and Tracking Methods”,
International Journal of Engineering Research &
Technology (IJERT), Vol. 3 Issue 11, November-2014.
[14] Junyan Peng, Patrick P. K. Chan, “Revised Naive Bayes
Classifier For Combating The Focus Attack In Spam
Filtering”, Proceedings of the 2013 International
Conference on Machine Learning and Cybernetics,
Tianjin, 14-17 July, 2013.
[15] KAVITHA.A and K.THULASIMANI, “Object
Tracking And Recognition For Real Time Videos Using
Naive Bayes Classification”, International Journal of
Computer Science and Mobile Applications, Vol.2
Issue. 1, January- 2014, pg. 63-78.
[16] http://www.openvisor.org R. Vzzani, R. Cucchiara,
"Video Surveillance Online Repository (ViSOR): an
Naïve based Moving Object Detection and Classification in Video using Background Subtraction
(IJSRD/Vol. 3/Issue 04/2015/426)
All rights reserved by www.ijsrd.com 1830
integrated framework" in Multimedia Tools and
Applications, vol. 50, n. 2, Kluwer Academic Press, pp.
359-380, 2010.
[17] http://i21www.ira.uka.de/image_sequences/ Karl-
Wilhelm-Straße: recorded by Klaus Fleischer, Michael
Haag, Frank Heimes, and Sven Noltemeier on
23.07.1997
[18] http://i21www.ira.uka.de/image_sequences/ rheinhafen
[19] http://i21www.ira.uka.de/image_sequences/ Karl-
Wilhelm-Straße: stationary camera installation by
German Eichberger, Klaus Fleischer and Holger Leuck.